Editing Openai/691c1dba-9228-800f-8463-13b3a9006306 (section)

=== simple example uses a local small summarizer (you can call your controller API) ===
def call_local_summarizer(text, max_tokens=400):
    # Placeholder: replace with HTTP call to your controller or local llama.cpp runner
    # For now, return a naive truncation summary
    return text[:400] + ("…" if len(text)>400 else "")

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("indir")
    args = ap.parse_args()
    p = Path(args.indir)
    outdir = p / "summaries"
    outdir.mkdir(exist_ok=True)
    for f in p.glob("*.json"):
        obj = json.loads(f.read_text(encoding='utf-8'))
        # canonical summary content candidates: hplm, hplm.summary, body fields
        text = ""
        if 'hplm' in obj:
            text = json.dumps(obj['hplm'], ensure_ascii=False)
        else:
            text = json.dumps(obj, ensure_ascii=False)
        summary = call_local_summarizer(text, max_tokens=400)
        outfile = outdir / (f.stem + "_summary.json")
        outobj = {"src": str(f), "summary": summary}
        outfile.write_text(json.dumps(outobj, ensure_ascii=False, indent=2), encoding='utf-8')
        print("Wrote", outfile)

if __name__=='__main__':
    main()

* Replace call_local_summarizer with your webui POST to a model you want to use for summarization (controller API). Keep summaries short (≤ 600 tokens).

==== 3) Compare & score (semantic similarity + evaluator) ====

We’ll produce a simple embedding-based similarity using sentence-transformers if you have it, or fallback to token-overlap.

Save compare_score.py:

python
#!/usr/bin/env python3